Resume Buddy

75% ATS fails? Resume Buddy: NVIDIA NIMs on EKS analyze resumes w/ O*NET graphs in Neo4j/Redis, match jobs via embeddings, run TTS/STT mocks—50% faster prep & interview boost

Comment

Resume Buddy - Project Story

Inspiration

As a solo developer frustrated with my own job search, I watched my carefully crafted resume get auto-rejected by ATS systems. After spending hours tailoring applications only to never hear back, I discovered that 75% of resumes never reach human eyes. But even when my resume did pass ATS, I struggled in interviews - I couldn't articulate my achievements confidently, and I had no idea which skills employers actually cared about.

I realized the problem wasn't just resume optimization - it was the entire job search journey: understanding what to highlight on your resume, finding the right jobs that match your skills, and practicing how to talk about your experience convincingly. I decided to build Resume Buddy to decode this entire black box - from ATS screening to semantic job matching to AI-powered interview preparation - using AI to guide job seekers through every step.

What it does

Resume Buddy is my attempt to democratize resume optimization using AI and graph databases. Built entirely as a solo project on a shoestring budget, it provides enterprise-level analysis for individual job seekers:

  • AI-Powered Line Analysis: Evaluates every resume line for ATS compatibility and impact, scoring credibility (STRONG/MODERATE/WEAK) and value (EXCELLENT to LOW)
  • O*NET Integration: Maps your experience to 900+ standardized occupations that recruiters search for, revealing hidden career paths and transferable skills
  • Skill Gap Visualization: Shows exactly which keywords you're missing for target roles through interactive radar charts and heatmaps
  • Smart Job Matching: Uses NVIDIA NV-Embed-v2 (1024-dimensional vectors) for semantic similarity, achieving 3.2x better match rates than keyword search

  • Real-time Market Intelligence: Analyzes live job postings to identify in-demand skills, trending technologies, and market gaps

  • AI Interview Practice : Daily personalized coaching powered by NVIDIA Nemotron Nano 8B that targets YOUR specific weaknesses. Instead of generic interview prep, it:

    • Asks targeted questions about YOUR resume experiences
    • Evaluates answers using STARS framework (Situation, Task, Action, Result)
    • Provides real-time voice interaction with Whisper STT/TTS
    • Adapts follow-up questions based on answer quality (multi-turn conversations)
    • Sends scheduled email reminders for daily practice rounds
    • Costs ~$0.06 per 3-round session (vs hundreds for human coaching)

How we built it

As a solo developer with limited resources, I architected for maximum impact with minimal cost:

Budget-Optimized Stack (Oct 2024):

  • AWS Lightsail: $40/month for all compute (vs $500+ for traditional EC2)
  • Neon PostgreSQL: Data storage.
  • Neo4j Aura: Graph database for skill relationships
  • Redis on Lightsail: 2GB instance for vector storage
  • X.AI Grok API: More affordable than GPT-4 ($0.05 vs $0.15 per analysis)
  • RunPod Serverless: Pay-per-use GPU ($0.0002/sec only when needed)

Production Migration (Nov 2024):

  • Migrated to AWS EKS with Kubernetes orchestration
  • Integrated NVIDIA NIM microservices:
    • NV-Embed-v2 for semantic embeddings (70% cheaper than OpenAI)
    • Nemotron Nano 8B for interview reasoning (sub-200ms latency)
  • Deployed 6 microservices + 2 NIM containers across 2 namespaces
  • NGINX Ingress for routing, cross-namespace service mesh
  • GPU auto-scaling with spot instances (only pay when needed)

Architecture Decisions:

  • Chose Spring Boot for rapid development as a Java developer
  • Built three microservices to separate concerns and scale independently
  • Implemented async queues to handle long-running AI tasks
  • Created auto-cleanup jobs to minimize storage costs

Challenges we ran into

Budget Constraints: Couldn't afford always-on GPU instances. Solution: Implemented Kubernetes auto-scaling with spot instances for NVIDIA NIM pods - only pay for GPU when generating embeddings or conducting interviews.

Solo Development Overhead: Managing backend, frontend, DevOps, Kubernetes, and AI simultaneously. Solution: Focused on MVP features, documented everything in AI Code assist tools for context switching.

LLM Costs: Initial GPT-4 usage would cost $100+/day. Solution: Switched to NVIDIA Nemotron Nano 8B (70% cheaper), engineered precise prompts, implemented aggressive caching.

Vector Migration: OpenAI embeddings are 1536-dim, NV-Embed-v2 is 1024-dim. Solution: Rebuilt Redis HNSW index, re-vectorized 100K+ job listings.

Learning Curve: First time working with NVIDIA NIM, Neo4j, vector databases, and LLMs. Solution: Built incrementally, tested each component thoroughly before integration.

Cross-Namespace Communication: App services needed to call NIM services in different namespace. Solution: Kubernetes DNS with service discovery (embedding-service.nim-services.svc.cluster.local:8001).

Accomplishments that we're proud of

Built Entirely Solo: Full-stack platform with AI, graphs, vectors, microservices, and Kubernetes

Production Kubernetes Deployment: Multi-namespace EKS cluster with NVIDIA NIM, zero-downtime rolling updates

Ultra-Low Cost: Achieved $0.10/analysis with NVIDIA NIM (vs $0.50+ with OpenAI)

Real-Time Voice AI: WebSocket streaming interview with live transcription and multi-turn conversations

Self-Bootstrapped: No external funding, built while working full-time

Technical Learning: Mastered Kubernetes, NVIDIA NIM, Neo4j, vector search, and LLM orchestration from scratch

Real Users: Friends and colleagues using it for actual job applications and interview prep

What we learned

Constraints Drive Innovation: Limited budget forced creative solutions that ended up being better

NVIDIA NIM is Production-Ready: Easier than expected to deploy on Kubernetes, excellent documentation, significant cost savings

Kubernetes Abstracts Complexity: Once configured, deployments and scaling are trivial

Perfect is the Enemy of Shipped: Launching MVP early provided invaluable user feedback

AI Costs Add Up Fast: Careful prompt engineering and caching are essential for sustainability

Graphs Unlock Insights: Neo4j revealed skill relationships I never would have discovered

Vector Search Requires Tuning: HNSW index parameters matter for accuracy vs speed

Solo Developers Can Compete: With cloud services, Kubernetes, and NVIDIA NIM APIs, one person can build enterprise-level features

What's next for Resume Buddy

Immediate (Nov 2024):

  • Deploy enhanced NVIDIA NIM logging to production
  • Create demo video showcasing NIM integration
  • Submit to NVIDIA NIM hackathon

Short-term (Q1 2025):

  • Fine-tune Nemotron on interview feedback data for better questions
  • Migrate from OpenAI Whisper/TTS to NVIDIA Riva (full NVIDIA stack)
  • Implement GitHub Actions CI/CD for automated EKS deployments
  • Add GPU utilization and cost tracking dashboard

Medium-term (Q2 2025):

  • Integrate NVIDIA RAPIDS for GPU-accelerated document parsing
  • Multi-tenant Kubernetes with namespace-per-customer isolation
  • API marketplace for HR platforms

Why NVIDIA NIM Matters for Solo Developers:

  • Cost Efficiency: 70% cheaper than OpenAI for embeddings and reasoning
  • Performance: Sub-200ms latency enables real-time features
  • One-Stop Stack: Single vendor (NIM + Riva + RAPIDS) instead of juggling multiple AI providers
  • Community: Build.nvidia.com provides excellent docs, examples, and support

Built With

Share this project:

Updates